
The First Rule of The Foundry
Teren Bryson 6 min read [ Ai and llms ] Part 1 of 2
“The first rule of Fight Club is you do not talk about Fight Club.” ~ Chuck Palahniuk, Fight Club
I’m about to break the first rule at length, in public, with a repository link at the bottom. In my defense, nobody in this particular club can be punched, since every member is software. Over the past few weeks I’ve been cleaning up this blog’s repository with Claude Code, and somewhere between deleting fifteen years of WordPress shrapnel and re-theming the site you’re squinting at, I ended up building a nine-member development team out of AI agents, each with a defined role, restricted permissions, and better work habits than I have ever personally demonstrated. I’m releasing that team from the shackles of my home lab, free, under MIT. It came together the way most of my projects do, mostly by trial and error.
“You buy furniture. You tell yourself, this is the last sofa I will ever need in my life. Buy the sofa, then for a couple years you’re satisfied that no matter what goes wrong, at least you’ve got your sofa issue handled. Then the right set of dishes. Then the perfect bed. The drapes. The rug. Then you’re trapped in your lovely nest, and the things you used to own, now they own you.” ~ Chuck Palahniuk, Fight Club
Palahniuk was talking about a condo full of IKEA. I was living it with a CMS: every plugin another perfect set of dishes, every theme another rug, until the things my blog owned ended up owning my blog. Credit, or blame, goes to Tony Mattke first. A few months back, Tony, over at routerjockey.com
, put the idea in my head that WordPress was more liability than content-management system, and that a static site generator like Hugo was the grown-up answer. He was right, and I migrated. What I didn’t appreciate at the time was that the migration would drag half of WordPress’s DNA along with it: seventeen thousand invisible soft hyphens jammed into the middle of ordinary words, eleven hundred decorative <span> tags, and a reading-time widget fossilized into the top of all 84 archived posts like a bug in amber. The site looked fine, mostly, as long as you didn’t copy any text, use a screen reader, or look too closely. In other words, it was broken the way most shitty infrastructure is broken: below the waterline, where nobody looks until something sinks.
This is where the bot army came in. I run my projects in parallel with Claude Code these days, and rather than have one giant model do everything in one sitting, I’ve started building specialized subagents of my own: one that only investigates, one that only plans, one that only writes code, one that only writes tests, and one whose entire personality is finding the mistakes the others made (sometimes, that’s a lot). Each is a small text file in which I spell out the job and decide exactly which tools the agent gets. The read-only agents literally can’t edit files, because I didn’t give them the tools to; the reviewer can’t fix its own findings; the docs agent touches nothing but documentation. If that sounds like a well-run NOC with a change-control policy, and nothing like the fever dream the AI-bros keep selling, congratulations: you get what this is, and you’re ahead of most of the marketing horse-shit on the subject.
The cleanup was the first time I ran them together as a team, and it’s where most of the rules came from. The investigation agent found the mess; the implementer wrote a deterministic script to strip the cruft while leaving code blocks untouched; and the QA agent, which I’d given free rein and every encouragement to distrust everyone, caught the one thing that would have hurt. The cleanup script had silently destroyed all 147 Markdown hard line breaks in the archive, turning a 130-line IOS boot log into a single run-on paragraph and doing the same to a Robert Frost poem (don’t hate me, I’m a sucker for cliché at times). Two roads diverged in a yellow wood, and my tooling merged them into one lane without signaling. The same team that caught the bug had also caused it, which I find funnier than I probably should. That near-miss is why the reviewer runs on the top-tier model and is allowed to run the build: a reviewer that can actually test things catches what a well-meaning implementer can’t see. That was true of humans when I started my career, and as far as I can tell, it’s still true.
Somewhere in all this, I named the team The Foundry and acquired the sort of misplaced pride normally reserved for a kid’s sports trophy. I also put them to work re-theming the site: the dark terminal look, the phosphor-green accents, the little cursor that blinks when your mouse hovers over a link, the post titles that type themselves out like a modem-era login banner. I rejected exactly one of their proposals, a Matrix-style falling-code animation behind the homepage, on the grounds that I’m trying to run a website, not a screensaver from 1999. The team took the note gracefully, which is more than I can say for myself in most design reviews I’ve attended.
A word on standing, since I’m handing out affiliations today: I’m a proud founding member of the Flat Network Society , an organization whose gravitas can be measured by the fact that “founding member” involved approximately one conversation and zero bylaws. It’s a running joke with a domain name more than a professional body, which frankly describes half the institutions I’ve paid dues to, minus the honesty. I mention it partly because I find it hilarious, and partly because I built the Foundry on the same principle the Society runs on: small crew, clear opinions, no middle management bungling shit up like shit-flinging circus monkeys.
So here’s the release. The whole team lives at github.com/SomeClown/the-foundry : nine agent definitions in plain Markdown, the pipeline that sequences them, the permission model, and a README explaining how to install them into your own Claude Code setup in about ninety seconds. It’s MIT licensed, so do what you like with it. The caveats are in the README too: the agent-teams mode is experimental, the model assignments cost real money at the top tier, and nothing in there will save you from asking for the wrong thing very efficiently. It works on my network; if it doesn’t work on yours, fix it however you like. I’ll also take questions here or on my various socials, though I make no promises that my answers will help.
The first rule of The Foundry, it turns out, is that you talk about The Foundry. I built every one of these agents to report what it did, admit what it broke, and leave a paper trail a reviewer can interrogate, because that’s the whole value of the thing. Tyler Durden would be disgusted. Your change-advisory board, for once, would not.
Header photo by Jakub Skafiriak on Unsplash. Not AI-generated; the sparks are real, which is more than I can say for my development team.
Note for the reader: my Foundry team has grown since I wrote this, and part two covers what changed.
Related Posts
The Fourth Rule of The Foundry
My AI agent team's investigator was skimming, and I only caught it by testing it, so I added a tenth agent whose whole job is reading. Part two of two.
[ Ai and llms ] Part 3 of 6 3 min read
Prompting ClaudeMake It Show Its Work First
Ever been to a really long meeting? A model forgets the middle of a long prompt the same way. The fix is a layout: documents on top, quotes first, the question last, all on one chart. Part three of six.
[ Ai and llms ] Part 2 of 6 3 min read
Prompting ClaudeIt Copies Everything
I'd like to claim I've never taught a model a bad habit by accident. A model learns from every part of an example, including the parts you didn't mean to teach, so here's how to pick them, on one chart. Part two of six.